A News Agent for What You Miss: The Bubble Is Small, the Avoidance Is Not

The problem

In June 2026, 42% of people across the markets surveyed by the Reuters Institute said they sometimes or often avoid the news, essentially unchanged on the year before and up from 29% in 2017, when the measure began (survey, self-reported; 97,520 respondents across 48 markets)¹. In Bulgaria, Croatia, Greece and Turkey the figure is 60% or more¹. Interest in news has fallen from 59% to just under 46% since the height of the pandemic (survey, self-reported)¹.

At the same time, more of the news people do see arrives through feeds ranked by someone else's software. In 2026, 30% of people said social media and video networks were their main source of news, up from 22% five years earlier, and these networks were for the first time the most widely used way of reaching online news, used by 54% of respondents (survey, self-reported)¹. Weekly use of AI chatbots for news rose from 7% to 10% in a year (survey, self-reported)¹.

The common worry is that these feeds lock people into "filter bubbles", showing them only what they already agree with. The evidence for that worry is weaker than its reputation. A 2022 Reuters Institute review of the research found that the algorithmic selection used by search engines and social platforms generally leads to slightly more diverse news use, the opposite of what the filter bubble idea predicts, and that echo chambers are much less widespread than commonly assumed². In the UK, around 2% of people were estimated to be in a left-leaning online news echo chamber and around 5% in a right-leaning one; in every country studied, many more internet users consumed no online news regularly than lived in partisan echo chambers (survey-based estimates from 2020 data, more than three years old)². The United States was the outlier: the only one of seven countries studied where more than 10% relied only on partisan news sources².

So the problem has two parts, and they are not the same size. A small minority sees a one-sided news diet. A much larger share sees little news at all, or avoids it, and so misses stories that bear on their lives. A news agent that "shows you what you are missing" would have to be judged against the second problem as much as the first.

What already works

Recommenders built for diversity

Researchers have built news recommenders (the software that ranks which articles a reader is shown) that aim for diversity instead of predicted clicks. In a five-week Swiss study, 151 German-speaking users read news through a custom app drawing on six Swiss outlets; one group received diversity-optimised recommendations, one accuracy-optimised recommendations and one a chronological feed⁵. The diverse group reported higher tolerance for opposing views than the chronological group, especially among conservative users, while political knowledge and participation did not differ (controlled experiment, small sample; published 2022)⁵. The authors note that participants did not perceive the conditions as different, and that the small sample makes claims about depolarisation tentative⁵.

A 2026 preprint, not yet peer reviewed, reports a five-week study with 120 US readers on the POPROX research platform: an algorithmic nudge that balanced domestic and world news increased exposure and consumption diversity, while the effect of presentation nudges written by a language model varied (controlled user study; abstract only)⁶.

In plain terms: diversity can be engineered into a recommender, and readers will read some of what it offers. Whether that changes what they know or think is much less clear.

Blindspot feeds in production

A simpler version already runs commercially. Ground News publishes a "Blindspot" feed of "stories disproportionately covered by one side of the political spectrum" (operator-reported)⁸. It flags a story as a blindspot for the left when fewer than 10 left-leaning sources report it and right-leaning coverage is at least 33%, and it excludes stories where more than 35% of coverage comes from low-factuality sources (operator-reported)⁹. Bias ratings are set for whole publications, not individual articles, by averaging three outside rating organisations, and the analysis is "done in the context of the U.S. political system" (operator-reported)⁹. The full feed sits behind a subscription⁸.

This matters because it shows the core idea working with almost no AI: counting sources, applying a fixed rule and labelling outlets.

AI assistants that answer news questions

General-purpose AI assistants already answer news questions for a growing minority. Their record is poor. A study by 22 public broadcasters from 18 countries, coordinated by the EBU and the BBC, evaluated more than 3,000 answers from four widely used AI assistants and found that 45% had at least one significant issue; in 31% of cases references were missing or misleading, and 20% of answers contained major accuracy issues such as fabricated or outdated information (expert evaluation, no control group; October 2025)⁷.

For a reader, this means that an agent built on assistants performing like those tested would put a significant problem into nearly half of its answers, a serious flaw in a tool meant to leave people better informed.

Rules that open the door

The EU Digital Services Act, Regulation (EU) 2022/2065 of 19 October 2022, provides that users of very large online platforms should "enjoy alternative options which are not based on profiling" for the main parameters of their recommender systems, and that such choices should be directly accessible where recommendations are presented (official legal text, recital 94)¹⁰. That creates legal room for alternatives to engagement-ranked feeds inside each platform. It does not, in the text read, give an outside agent access to a user's feed.

What does not exist

No trial was found of a personal agent that sees a reader's actual news diet across platforms, judges which important stories and viewpoints are absent, and surfaces them. Three pieces would have to be invented or obtained: access to a person's reading across platforms, which no source reviewed showed being available to third parties; a credible, article-level way of deciding what is "important" and what counts as a different viewpoint outside the US left–right frame; and summarisation accurate enough to trust.

The possibility

One way this could work: a reader would install an agent that, with their consent, read their news history across the apps and sites they use. Once a day it would compare that diet with what a broad panel of outlets covered, and send a short note: three stories that were widely reported and bear on the reader's place, work or stated interests but that they did not see; and, for one story they did follow, a link to the strongest coverage from a different perspective. Each item would carry a plain reason ("covered by 40 outlets, none in your feed"), a link to the original reporting, and no rewritten summary unless the reader asked for one.

The more useful question for such an agent would be less "which side am I missing?" and more "which important story did I not see at all?"

The role of AI would be specific. Machine learning would be needed to group articles from different outlets and languages into the same story, to classify the stance or framing of individual articles instead of relying on outlet labels, and to judge what a given reader would find relevant enough to open. A generative model could write the one-line reason for each item. Agent behaviour, acting on the reader's behalf across services, would be needed only if the agent had to collect the reading history itself rather than receive it.

Much of the rest would not be AI at all. Access to reading history would be a question of data rights and platform policy. The panel of outlets would be an editorial choice. The business model, whether a subscription, a public-service offer or a publisher consortium, would be a question of money. And the decision about what counts as "important" would be a value judgement that software could apply but not settle.

If the evidence above holds, such an agent would do more good by targeting omission than by targeting slant: helping the large group who see little news find the few stories that matter to them, rather than mainly correcting the small group whose news is one-sided.

What would have to be true

1. Readers would have to want it. With 42% avoiding the news at least sometimes¹, an agent that pushed more news could be ignored by exactly the people who miss most. It would have to be brief and plainly relevant.
2. The agent would need to see the reader's diet. Without access to what a person actually reads across platforms, it could not know what they were missing. The non-profiling option in EU law gives users a choice inside each platform¹⁰, not a route for an outside agent.
3. "Important" would need a defensible, visible definition. Counting how widely a story is covered, as blindspot feeds do⁹, is transparent but crude; anything more sophisticated would need to show its reasoning.
4. Viewpoint labels would have to work outside the US and at article level. The production ratings found are set per publication and framed for US politics⁹, while the research shows the US is an outlier². This is an AI condition: classifying the framing of single articles across languages and party systems.
5. The AI parts would need to be far more reliable than today's general assistants. With 45% of tested answers carrying a significant issue⁷, the agent would have to link to original reporting rather than summarise it, or reach a much lower error rate.
6. Wider exposure would have to produce a benefit people can see. The evidence that it changes attitudes is weak³⁴⁵, so the case would rest on readers being better informed about specific events, which would itself need measuring.

What could stop it

What the evidence shows

Large changes to what people see have produced small changes in what they think. In a 2020 US experiment, moving consenting Facebook and Instagram users to chronological feeds substantially reduced their time on the platforms and increased the political and untrustworthy content they saw, but did not significantly change issue polarisation, affective polarisation or political knowledge over three months (randomised field experiment; abstract)³. In a companion study of 23,377 Facebook users, the median user received 50.4% of content from like-minded sources and 14.7% from cross-cutting ones; cutting like-minded exposure by about a third had no measurable effect on eight preregistered attitude measures, including affective polarisation and belief in false claims (randomised field experiment, 2020)⁴.

Put simply: in these two large experiments, changing the feed changed the diet but not the views.

The diversity-recommender evidence is small and partly inconclusive: participants in the Swiss study did not perceive the diverse and narrow conditions as different⁵, and the 2026 study is a preprint read in abstract only⁶. AI answers about news remained error-prone across all countries and languages tested⁷. The one production blindspot feed found is subscription-based and relies on US-framed, publication-level ratings⁸⁹.

Reasoning

What would make this possibility less likely, less useful or more harmful than it first appears?

It may solve the smaller problem. If echo chambers are as rare outside the US as the research suggests², an agent sold on "breaking your bubble" would address a minority while missing the larger group who opt out of news altogether.

It would concentrate a sensitive power. An agent that decided what you "should" have seen would be making editorial judgements for everyone who used it. If its idea of balance were skewed, or shaped by whoever paid for it, it could narrow views while claiming to widen them. Who decides the outlet panel, and who audits it, would matter more than the model.

It would require intimate data. A full record of what someone reads is politically revealing. Holding it would create privacy and security risks that the benefit would have to justify.

Incentives would cut against it. Platforms that earn from engagement would have little reason to open their feeds to an outside agent. Publishers might welcome the traffic but not a new intermediary between them and their readers.

“Who decides the outlet panel, and who audits it, would matter more than the model.”

Simpler options may do most of the job. A rule-based blindspot list⁸, a public broadcaster's daily summary of what matters, or a platform's non-profiling feed¹⁰ each address part of the problem without the cost and risk of a personal agent. If they deliver most of the value, the AI version would add complexity for a small gain.

Signs to watch

• Further results from the POPROX research platform and similar testbeds, especially studies that measure what readers know, not only what they click⁶.
• How very large platforms present and report on their non-profiling recommender options under the Digital Services Act, and any enforcement decisions on recommender transparency¹⁰.
• The next annual Reuters Institute Digital News Report figures on news avoidance and on use of AI chatbots for news¹.
• Any repeat of the EBU–BBC evaluation showing whether AI assistants' error rates on news have fallen⁷.
• Launch of any service that lets people export their news reading history from platforms to a third-party tool, with consent.
• Whether blindspot-style features extend beyond US political categories to other countries' party systems⁹.